The dataset .json annotations are great for dataloading, however it would be really helpful to have more meaningful description of the entries. For example in LTA task, what is a difference between:
'interval_{start|end}_{sec|frame}'
and
'action_clip_{start|end}_{sec|frame}'
Or similarly, how are the data such as
['annotations'][idx]['noun']
formatted.
This is a dummy example, but most tasks have somewhat ambiguous annotations whose meaning is only accessible through trial and error and analysing benchmark codes/submissions. It would be very helpful to have these descriptions in the schema definitions.